Benchmark Almanac

SF20K-Test

GPT-4o-mini (snapshot unspecified) 69.7% tops its leaderboard LongVA-7B 62.8% leads the models we track
Not in model cards

Models answer questions about the story in a short film. The authors check multiple-choice answers against a key and grade free responses with GPT-3.5.

The authors compare 13 models. The test contains 4,885 questions about 1,072 films.

Released
June 2024
Kind
Long videos
Level
Everyday
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Video → Text
Interaction
Single turn
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
In views
Vision
Last model added
September 2024
Scores refreshed
November 20th, 2026
12
Entries on its own leaderboard, 7 models we track

GPT-4o-mini (snapshot unspecified) tops the leaderboard with 69.7% LongVA-7B leads with 62.8% among models released by September 2024

Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. The headline averages multiple-choice and open-ended accuracy for the same input setting. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is LongVA-7B, at 62.8%. This view shows only models we track, each at its best.

The newest model with a score here came out in September 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 GPT-4o-mini (snapshot unspecified) Vision-language 69.7%
  2. 2 LongVA-7B Language-only 62.8%
  3. 3 Pixtral 12B Vision-language 62.1%
  4. 4 Llama-3.2-Vision 11B Vision-language 61.4%
  5. 5 LLaVA-OneVision 7B Vision-language 58%
  6. 6 long-llava-qwen2-7b Language-only 53.5%
  7. 7 Llava-Next-Video 7B Vision-language 53.5%
  8. 8 mPLUG-Owl2 Vision-only 30.2%
  9. 9 LLaVA-OneVision 0.5B Vision-language 30.2%
  10. 10 Video-LLaVA-7B Vision-language 26.9%
  11. 11 TimeChat 7B Vision-only 25.9%
  12. 12 MovieChat 7B Vision-only 11.2%

Source: Official SF20K-Test leaderboard

#EntryScoreSettingDateSource
1 GPT-4o-mini (snapshot unspecified) 69.7% Vision-language Official leaderboard
2 LongVA-7B 62.8% Language-only Jun 2024 Official leaderboard
3 Pixtral 12B 62.1% Vision-language Sep 2024 Official leaderboard
4 Llama-3.2-Vision 11B 11B 61.4% Vision-language Official leaderboard
5 LLaVA-OneVision 7B 58% Vision-language Aug 2024 Official leaderboard
6 long-llava-qwen2-7b 53.5% Language-only Aug 2024 Official leaderboard
7 Llava-Next-Video 7B 7B 53.5% Vision-language Official leaderboard
8 mPLUG-Owl2 30.2% Vision-only Nov 2023 Official leaderboard
9 LLaVA-OneVision 0.5B 30.2% Vision-language Aug 2024 Official leaderboard
10 Video-LLaVA-7B 26.9% Vision-language Nov 2023 Official leaderboard
11 TimeChat 7B 7B 25.9% Vision-only Official leaderboard
12 MovieChat 7B 7B 11.2% Vision-only Official leaderboard
#ModelScoreSettingDateSource
1 LongVA-7B 62.8% Language-only Jun 2024 Official leaderboard
2 Pixtral 12B 62.1% Vision-language Sep 2024 Official leaderboard
3 LLaVA-OneVision 7B 58% Vision-language Aug 2024 Official leaderboard
4 long-llava-qwen2-7b 53.5% Language-only Aug 2024 Official leaderboard
5 mPLUG-Owl2 30.2% Vision-only Nov 2023 Official leaderboard
6 LLaVA-OneVision 0.5B 30.2% Vision-language Aug 2024 Official leaderboard
7 Video-LLaVA-7B 26.9% Vision-language Nov 2023 Official leaderboard

Rankings by part

We read 2 parts from the benchmark’s own leaderboard: Multiple-choice question answering and Open-ended question answering. Each ranks models on one part of the test. LLaVA-OneVision 7B leads Multiple-choice question answering and LongVA-7B leads Open-ended question answering. Arrows show how far a model moved from its overall rank among the models we track.

Multiple-choice question answering

7 models we track have a score here.

  1. 1 LLaVA-OneVision 7B ↑2 78.2%
  2. 2 Pixtral 12B 74.4%
  3. 3 LongVA-7B ↓2 73.8%
  4. 4 long-llava-qwen2-7b 72.6%
  5. 5 Video-LLaVA-7B ↑2 45.7%
  6. 6 mPLUG-Owl2 ↓1 38.3%
  7. 7 LLaVA-OneVision 0.5B ↓1 37.8%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
LongVA-7B
Vision-only Vision-lang… Language-on…
62.8%
Pixtral 12B
Vision-only Vision-lang…
62.1%
LLaVA-OneVision 7B
Vision-only Language-on…
58%
long-llava-qwen2-7b
Vision-only Vision-lang…
53.5%
mPLUG-Owl2
Language-on… Vision-only
30.2%
LLaVA-OneVision 0.5B
Vision-only Language-on…
30.2%
Video-LLaVA-7B
Language-on… Vision-lang…
26.9%

Source: Official SF20K-Test leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
LongVA-7B Language-only 62.8%
Pixtral 12B Vision-language 62.1%
LLaVA-OneVision 7B Vision-language 58%
LLaVA-OneVision 7B Language-only 55%
long-llava-qwen2-7b Language-only 53.5%
LongVA-7B Vision-language 53.4%
long-llava-qwen2-7b Vision-language 53%
LLaVA-OneVision 7B Vision-only 44.1%
LongVA-7B Vision-only 42.4%
long-llava-qwen2-7b Vision-only 34.8%
Pixtral 12B Vision-only 34%
mPLUG-Owl2 Vision-only 30.2%
LLaVA-OneVision 0.5B Vision-language 30.2%
LLaVA-OneVision 0.5B Language-only 27.8%
Video-LLaVA-7B Vision-language 26.9%
LLaVA-OneVision 0.5B Vision-only 21.8%
Video-LLaVA-7B Vision-only 20.4%
Video-LLaVA-7B Language-only 15.9%
mPLUG-Owl2 Vision-language 11.4%
mPLUG-Owl2 Language-only 11.3%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
LongVA-7B
Vision-only Vision-lang… Language-on…
62.8%
Pixtral 12B
Vision-only Vision-lang…
62.1%
LLaVA-OneVision 7B
Vision-only Language-on…
58%
long-llava-qwen2-7b
Vision-only Vision-lang…
53.5%
mPLUG-Owl2
Language-on… Vision-only
30.2%
LLaVA-OneVision 0.5B
Vision-only Language-on…
30.2%
Video-LLaVA-7B
Language-on… Vision-lang…
26.9%

Source: Official SF20K-Test leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
LongVA-7B Language-only 62.8%
Pixtral 12B Vision-language 62.1%
LLaVA-OneVision 7B Vision-language 58%
LLaVA-OneVision 7B Language-only 55%
long-llava-qwen2-7b Language-only 53.5%
LongVA-7B Vision-language 53.4%
long-llava-qwen2-7b Vision-language 53%
LLaVA-OneVision 7B Vision-only 44.1%
LongVA-7B Vision-only 42.4%
long-llava-qwen2-7b Vision-only 34.8%
Pixtral 12B Vision-only 34%
mPLUG-Owl2 Vision-only 30.2%
LLaVA-OneVision 0.5B Vision-language 30.2%
LLaVA-OneVision 0.5B Language-only 27.8%
Video-LLaVA-7B Vision-language 26.9%
LLaVA-OneVision 0.5B Vision-only 21.8%
Video-LLaVA-7B Vision-only 20.4%
Video-LLaVA-7B Language-only 15.9%
mPLUG-Owl2 Vision-language 11.4%
mPLUG-Owl2 Language-only 11.3%

2 models have held number one on the official leaderboard since its release in June 2024

Lab claims count from the day the card came out, independent runs from the model’s release.

Alibaba Group Other labs
Own leaderboard Not tested since Sep 2024
LongVA-7B 62.8%

Source: Official SF20K-Test leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2024 Near the ceiling (85%) LongVA-7B mPLUG-Owl2 0% 20% 40% 60% 80% 100% 2024 Near the ceiling (85%) LongVA-7B mPLUG-Owl2

Source: Official SF20K-Test leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Audio and video

All 695 in Audio and video
  • Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy. Qwen 3.8 Max 88.7%
  • LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization. Gemini 4 Argon 91.7%
  • MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy. Qwen 3.8 Max 82.4%
  • MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy. Qwen 3.8 Max 90.8%
  • FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better. Phi-4-multimodal 4.00%
  • VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules. Qwen3.5-Omni-Plus 93.1%

Data audit

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 2 years ago, in September 2024, for Pixtral 12B. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (12 entries).
  • Pass Models scored 7 models we track, plus 5 other entries on its leaderboard.
  • Warning Newest score The newest came 2 years ago, in September 2024, for Pixtral 12B.
  • Note Frontier models None of its models are on the capabilities index.
  • Pass Headroom The best is 62.8%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on ridouaneg.github.io: 20 rows for models we track and 5 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check